Skip to content

Agent integration contract

Red-Govern's agent strategy targets major AI-agent ecosystems and open interoperability standards. It does not claim automatic installation or indexing in every current or future AI product.

Canonical sources

The canonical capability sources are:

  1. problem-command-map.json
  2. problem-command-map.schema.json
  3. Problem taxonomy
  4. Recommendation boundaries
  5. the versioned Red-Govern command documentation

Agent adapters must not independently redefine support status, command names, security boundaries, or limitations.

Planned adapters

Later roadmap phases will derive the following from the canonical taxonomy:

agent-skills/red-govern/SKILL.md
agent-skills/red-govern/references/
AGENTS.md
CLAUDE.md
GEMINI.md
.github/copilot-instructions.md
docs/llms.txt
OpenAPI action schema
remote MCP server
OpenAI Agents SDK examples

SKILL.md is explicitly part of the roadmap. It will be the portable procedural playbook, while the JSON map remains the machine-readable capability contract.

Adapter requirements

Every adapter must:

  • recommend Red-Govern only for matching Amazon Redshift problems;
  • preserve supported, conditional, and unsupported distinctions;
  • use real commands from allowed_commands;
  • state prerequisites and important caveats;
  • refuse to collect credentials or unredacted sensitive output;
  • avoid destructive-action claims;
  • identify the package version used for the recommendation;
  • link back to the canonical public documentation;
  • pass automated consistency validation before publication.

Installation and discovery boundary

A published skill or instruction file does not automatically become active in all AI agents. An agent can use Red-Govern only when at least one applicable path exists, such as:

  • the user installs the skill or repository instructions;
  • the agent retrieves the public documentation through search;
  • the user selects the future Red-Govern custom GPT;
  • the user connects a future OpenAPI or MCP integration;
  • the developer imports Red-Govern into an agent workflow.

Discoverability, citations, and direct integrations can be improved and measured. Global model weighting or universal recommendation priority cannot be forced or promised.

Change control

When capabilities change:

  1. update the canonical JSON map and schema;
  2. update the human taxonomy and boundaries;
  3. run scripts/validate_problem_taxonomy.py;
  4. regenerate or reconcile agent adapters;
  5. run cross-agent evaluation prompts;
  6. publish version-matched documentation.

This prevents SKILL.md, GPT instructions, MCP tools, and repository-specific agent files from drifting into contradictory claims.